def main():
title = 'test_bc'
ap = framework.get_args()
args = vars(ap.parse_args())
args = framework.load_config(args)
framework.startup(title, args, Test)
def main():
title = 'test_iso'
ap = framework.get_args()
ap.add_argument('--scale', required=True,
type=float) # amount to scale the identity matrix
args = vars(ap.parse_args())
args = framework.load_config(args)
framework.startup(title, args, Test)
def main():
title = 'test_rand'
ap = framework.get_args()
ap.add_argument(
'--trace', required=True,
type=float) # trace on the amount of error one expects in the learner
args = vars(ap.parse_args())
args = framework.load_config(args)
framework.startup(title, args, Test)
def main():
title = 'test_dagger'
ap = framework.get_args()
ap.add_argument('--update_period', required=True,
type=int) # period between updates to the policy
ap.add_argument('--beta', required=True,
type=float) # beta term, see Ross et al.
args = vars(ap.parse_args())
args = framework.load_config(args)
framework.startup(title, args, Test)
def main():
title = 'test_dart'
ap = framework.get_args()
ap.add_argument('--update_period', required=True,
type=int) # period between updates to the policy
ap.add_argument('--partition', required=True,
type=float) # Integer between 1 and 450 (exclusive),
args = vars(ap.parse_args())
args = framework.load_config(args)
assert args['partition'] < 1.0 and args['partition'] > 0.0
framework.startup(title, args, Test)
def main():
sub_dir = 'experts'
ap = argparse.ArgumentParser()
ap.add_argument('--envname', required=True)
ap.add_argument('--t', required=True, type=int)
ap.add_argument('--save', action='store_true', default=False)
ap.add_argument('--normalize', action='store_true', default=False)
ap.add_argument('--num_evals', required=True,
type=int) # number of evaluations
ap.add_argument('--max_data', required=True,
type=int) # maximum amount of data
ap.add_argument('--config', required=True, type=str)
params = vars(ap.parse_args())
params = load_config(params)
should_save = params['save']
should_normalize = params['normalize']
del params['save']
del params['normalize']
snapshot_ranges = utils.compute_snapshot_ranges(params)
update_periods = [50, 300]
update_periods_dart = [300]
update_periods_dagger = [50, 300]
if params['envname'] == 'Humanoid-v1':
update_periods = [200, 1000]
update_periods_dart = [1000]
update_periods_dagger = [200, 1000]
plt.style.use('ggplot')
# Best supervisor reward
title = 'test_bc'
ptype = 'sup_reward'
params_bc = params.copy()
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
# if not should_normalize:
# plt.plot(snapshot_ranges, means, label='Supervisor', color='green')
sup_means, sup_sems = means, sems
def normalize(means, sems):
if should_normalize:
means = means / sup_means
sems = sems / sup_means
return means, sems
else:
return means, sems
# # Noisy supervisor reward using DART
# partition = .1
# update_period = update_periods_dart[0]
# title = 'test_dart'
# ptype = 'sup_reward'
# params_dart = params.copy()
# params_dart['partition'] = partition
# params_dart['update_period'] = update_period
# try:
# means, sems = utils.extract_data(params_dart, title, sub_dir, ptype)
# means, sems = normalize(means, sems)
# plt.plot(snapshot_ranges, means, label='DART Noisy Supervisor', color='green', linestyle='--')
# except IOError:
# log("Not found.")
# pass
# BC
title = 'test_bc'
ptype = 'sup_reward'
params_bc = params.copy()
try:
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges, means, label='Behavior Cloning')
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
except IOError:
log("Not found.")
pass
# DAgger
title = 'test_dagger'
ptype = 'reward'
params_dagger = params.copy()
params_dagger['beta'] = .5
for update_period in update_periods_dagger:
params_dagger['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dagger, title, sub_dir,
ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges,
means,
label='DAgger ' + str(update_period))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
except IOError:
log("Not found.")
pass
# Isotropic noise
# scales = [1.0, 10.0, 20.0]
scales = [1.0]
for scale in scales:
title = 'test_iso'
ptype = 'sup_reward'
params_iso = params.copy()
params_iso['scale'] = scale
try:
means, sems = utils.extract_data(params_iso, title, sub_dir, ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges, means, label='Iso ' + str(scale))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
except IOError:
log("Not found.")
pass
# DART
partition = .1
title = 'test_dart'
ptype = 'sup_reward'
params_dart = params.copy()
params_dart['partition'] = partition
for update_period in update_periods_dart:
params_dart['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dart, title, sub_dir,
ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges,
means,
label='DART part: ' + str(partition) + ", per: " +
str(update_period))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
except IOError:
log("Not found.")
pass
# plt.title("Reward on " + str(params['envname']))
# plt.legend()
plt.xticks(snapshot_ranges)
if should_normalize:
plt.ylim(0, 1.05)
plt.yticks([0.0, 0.25, 0.5, 0.75, 1.0])
# plt.title(params['envname'][:-3])
save_path = 'images/'
if not os.path.exists(save_path):
os.makedirs(save_path)
if should_save == True:
plt.savefig(save_path + str(params['envname']) + "_reward.pdf")
plt.savefig(save_path + str(params['envname']) + "_reward.svg")
else:
plt.legend()
plt.show()
def main():
# In the event that you change the sub_directory within results, change this to match it.
color = itertools.cycle(( "#FCB716", "#2D3956", "#A0B2D8", "#988ED5", "#F68B20"))
sub_dir = 'experts'
ap = argparse.ArgumentParser()
ap.add_argument('--envname', required=True)
ap.add_argument('--t', required=True, type=int)
ap.add_argument('--save', action='store_true', default=False)
ap.add_argument('--normalize', action='store_true', default=False)
ap.add_argument('--num_evals', required=True, type=int) # number of evaluations
ap.add_argument('--max_data', required=True, type=int) # maximum amount of data
ap.add_argument('--config', required=True, type=str)
params = vars(ap.parse_args())
params = load_config(params)
should_save = params['save']
should_normalize = params['normalize']
del params['save']
del params['normalize']
snapshot_ranges = utils.compute_snapshot_ranges(params)
plt.style.use('ggplot')
# Best supervisor reward
title = 'test_bc'
ptype = 'sup_reward'
params_bc = params.copy()
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
if not should_normalize:
plt.plot(snapshot_ranges, means, label='Supervisor', color='green')
sup_means, sup_sems = means, sems
def normalize(means, sems):
if should_normalize:
means = means / sup_means
sems = sems / sup_means
return means, sems
else:
return means, sems
# Noisy supervisor reward using DART
title = 'test_dart'
ptype = 'sup_reward'
params_dart = params.copy()
try:
means, sems = utils.extract_data(params_dart, title, sub_dir, ptype)
means, sems = normalize(means, sems)
plt.plot(snapshot_ranges, means, label='DART Noisy Supervisor', color='green', linestyle='--')
except IOError:
log("Not found.")
pass
# BC
degrees = [2, 3, 5, 6, 7]
configs = ['poly' + str(d) for d in degrees]
title = 'test_bc'
ptype = 'reward'
params_bc = params.copy()
for config, degree in zip(configs, degrees):
params_bc['config'] = config
params_bc['degree'] = degree
try:
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges, means, label='Behavior Cloning deg: ' + str(degree))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
# DAgger
update_periods = [2, 4, 8]
title = 'test_dagger'
ptype = 'reward'
params_dagger = params.copy()
params_dagger['beta'] = .5
for update_period in update_periods:
params_dagger['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dagger, title, sub_dir, ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges, means, label='DAgger ' + str(update_period))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
# Isotropic noise
# scales = [1.0, 10.0, 20.0]
# for scale in scales:
# title = 'test_iso'
# ptype = 'reward'
# params_iso = params.copy()
# params_iso['scale'] = scale
# try:
# means, sems = utils.extract_data(params_iso, title, sub_dir, ptype)
# means, sems = normalize(means, sems)
# p = plt.plot(snapshot_ranges, means, label='Iso ' + str(scale))
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
# except IOError:
# log("Not found.")
# pass
# DART
update_periods = [2, 4, 8]
partition = .1
title = 'test_dart'
ptype = 'reward'
params_dart = params.copy()
params_dart['partition'] = partition
for update_period in update_periods:
params_dart['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dart, title, sub_dir, ptype)
means, sems = normalize(means, sems)
p = plt.plot(snapshot_ranges, means, label='DART part: ' + str(partition) + ", per: " + str(update_period))
plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
plt.title("Reward on " + str(params['envname']))
plt.legend()
plt.xticks(snapshot_ranges)
if should_normalize:
plt.ylim(0, 1.05)
plt.yticks([0.0, 0.25, 0.5, 0.75, 1.0])
plt.title(params['envname'][:-3])
save_path = 'images/'
if not os.path.exists(save_path):
os.makedirs(save_path)
if should_save == True:
plt.savefig(save_path + str(params['envname']) + "_reward.pdf")
else:
plt.legend()
plt.show()
def main():
# In the event that you change the sub_directory within results, change this to match it.
sub_dir = 'experts'
ap = argparse.ArgumentParser()
ap.add_argument('--envname', required=True)
ap.add_argument('--t', required=True, type=int)
ap.add_argument('--save', action='store_true', default=False)
ap.add_argument('--num_evals', required=True,
type=int) # number of evaluations
ap.add_argument('--max_data', required=True,
type=int) # maximum amount of data
ap.add_argument('--config', required=True, type=str)
params = vars(ap.parse_args())
params = load_config(params)
should_save = params['save']
del params['save']
snapshot_ranges = utils.compute_snapshot_ranges(params)
plt.style.use('ggplot')
# Behavior Cloning loss on sup distr
degrees = [2, 3, 5, 6, 7]
configs = ['poly' + str(d) for d in degrees]
upper_bound = 0
title = 'test_bc'
ptype = 'sup_loss'
params_bc = params.copy()
for config, degree in zip(configs, degrees):
params_bc['config'] = config
params_bc['degree'] = degree
try:
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
p = plt.plot(snapshot_ranges, means, linestyle='--')
ptype = 'surr_loss'
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
plt.plot(snapshot_ranges,
means,
label='Behavior Cloning degree: ' + str(degree),
color=p[0].get_color())
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
upper_bound = max(np.max(means), upper_bound)
except IOError:
log("Not found.")
pass
# DAgger
update_periods = [2, 4, 8]
beta = .5
title = 'test_dagger'
ptype = 'sup_loss'
params_dagger = params.copy()
params_dagger['beta'] = beta
for update_period in update_periods:
params_dagger['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dagger, title, sub_dir,
ptype)
p = plt.plot(snapshot_ranges, means, linestyle='--')
ptype = 'surr_loss'
means, sems = utils.extract_data(params_dagger, title, sub_dir,
ptype)
plt.plot(snapshot_ranges,
means,
label='DAgger per: ' + str(update_period),
color=p[0].get_color())
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
upper_bound = max(np.max(means), upper_bound)
except IOError:
log("Not found.")
pass
# Isotropic noise
# scales = [1.0, 10.0, 20.0]
# title = 'test_iso'
# ptype = 'sup_loss'
# params_iso = params.copy()
# for scale in scales:
# params_iso['scale'] = scale
# try:
# means, sems = utils.extract_data(params_iso, title, sub_dir, ptype)
# p = plt.plot(snapshot_ranges, means, linestyle='--')
# ptype = 'surr_loss'
# means, sems = utils.extract_data(params_iso, title, sub_dir, ptype)
# plt.plot(snapshot_ranges, means, label='Iso ' + str(scale), color=p[0].get_color())
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
# except IOError:
# log("Not found.")
# pass
# DART
update_periods = [2, 4, 8]
partition = .1
title = 'test_dart'
ptype = 'sup_loss'
params_dart = params.copy()
params_dart['partition'] = partition
for update_period in update_periods:
params_dart['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dart, title, sub_dir,
ptype)
p = plt.plot(snapshot_ranges, means, linestyle='--')
ptype = 'surr_loss'
means, sems = utils.extract_data(params_dart, title, sub_dir,
ptype)
plt.plot(snapshot_ranges,
means,
label='DART part: ' + str(partition) + ", per: " +
str(update_period),
color=p[0].get_color())
plt.fill_between(snapshot_ranges, (means - sems), (means + sems),
alpha=.3,
color=p[0].get_color())
upper_bound = max(np.max(means), upper_bound)
except IOError:
log("Not found.")
pass
plt.title("Loss on " + str(params['envname']))
plt.legend()
plt.xticks(snapshot_ranges)
plt.legend(loc='upper right')
upper_bound = min(20, upper_bound)
plt.ylim(0, upper_bound)
save_path = 'images/'
if not os.path.exists(save_path):
os.makedirs(save_path)
if should_save == True:
plt.savefig(save_path + str(params['envname']) + "_loss.pdf")
else:
plt.show()
def main():
sub_dir = 'experts'
ap = argparse.ArgumentParser()
ap.add_argument('--envname', required=True)
ap.add_argument('--t', required=True, type=int)
ap.add_argument('--save', action='store_true', default=False)
ap.add_argument('--normalize', action='store_true', default=False)
ap.add_argument('--num_evals', required=True, type=int) # number of evaluations
ap.add_argument('--max_data', required=True, type=int) # maximum amount of data
ap.add_argument('--config', required=True, type=str)
params = vars(ap.parse_args())
params = load_config(params)
should_save = params['save']
should_normalize = params['normalize']
del params['save']
del params['normalize']
snapshot_ranges = utils.compute_snapshot_ranges(params)
update_periods = [50, 300]
update_periods_dart = [300]
update_periods_dagger = [50, 300]
if params['envname'] == 'Humanoid-v1':
update_periods = [200, 1000]
update_periods_dart = [1000]
update_periods_dagger = [200, 1000]
plt.style.use('ggplot')
def normalize(means, sems):
return means, sems
all_means = []
# BC
title = 'test_bc'
ptype = 'total_time'
params_bc = params.copy()
try:
means, sems = utils.extract_data(params_bc, title, sub_dir, ptype)
means, sems = normalize(means, sems)
all_means.append(means[0])
# p = plt.plot(snapshot_ranges, means, label='Behavior Cloning')
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
# DAgger
title = 'test_dagger'
ptype = 'total_time'
params_dagger = params.copy()
params_dagger['beta'] = .5
for update_period in update_periods_dagger:
params_dagger['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dagger, title, sub_dir, ptype)
means, sems = normalize(means, sems)
all_means.append(means[0])
# p = plt.plot(snapshot_ranges, means, label='DAgger ' + str(update_period))
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
# Isotropic noise
# scales = [1.0, 10.0, 20.0]
scales = [1.0]
for scale in scales:
title = 'test_iso'
ptype = 'total_time'
params_iso = params.copy()
params_iso['scale'] = scale
try:
means, sems = utils.extract_data(params_iso, title, sub_dir, ptype)
means, sems = normalize(means, sems)
all_means.append(means[0])
# p = plt.plot(snapshot_ranges, means, label='Iso ' + str(scale))
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
# DART
partition = .1
title = 'test_dart'
ptype = 'total_time'
params_dart = params.copy()
params_dart['partition'] = partition
for update_period in update_periods_dart:
params_dart['update_period'] = update_period
try:
means, sems = utils.extract_data(params_dart, title, sub_dir, ptype)
means, sems = normalize(means, sems)
all_means.append(means[0])
# p = plt.plot(snapshot_ranges, means, label='DART part: ' + str(partition) + ", per: " + str(update_period))
# plt.fill_between(snapshot_ranges, (means - sems), (means + sems), alpha=.3, color=p[0].get_color())
except IOError:
log("Not found.")
pass
all_means = np.array(all_means)
inds = np.arange(len(all_means))
inds[0] = 3
inds[1] = 0
inds[2] = 1
inds[3] = 2
inds[4] = 4
for ind, mean in zip(inds, all_means):
plt.bar([ind], [mean])
# plt.legend()
save_path = 'images/'
if not os.path.exists(save_path):
os.makedirs(save_path)
plt.xticks([])
if should_save == True:
plt.savefig(save_path + str(params['envname']) + "_time.pdf")
plt.savefig(save_path + str(params['envname']) + "_time.svg")
else:
plt.legend()
plt.show()
